Predicting Cellular Responses to Perturbation Across Diverse Contexts with State

Learn about State, a transformer model designed to predict cellular responses to perturbations across diverse biological contexts in this conference talk. Discover how this machine learning approach addresses the challenge of generalizing perturbation effects to unobserved cellular contexts by accounting for cellular heterogeneity within and across experiments. Explore the model's training on gene expression data from over 100 million perturbed cells and its ability to predict perturbation effects across sets of cells. Examine the significant performance improvements, including more than 30% better discrimination of effects on large datasets and enhanced accuracy in identifying differentially expressed genes across genetic, signaling, and chemical perturbations. Understand how State's cell embedding, trained on observational data from 167 million cells, enables identification of strong perturbations in novel cellular contexts without prior training exposure. Learn about Cell-Eval, a comprehensive evaluation framework that demonstrates the model's capability to detect cell type-specific perturbation responses, including cell survival mechanisms. Gain insights into how this research advances virtual cell model development and its implications for understanding biological mechanisms and drug target selection in the field of AI for drug discovery.